IC-903Knowledge editing performance (ES, GS, LS) improves as model scale increases from GPT-2 (124M) to T5-XL (2.8B) to GPT-J (6B) across all editing methods

Chenmien Tan, Ge Zhang, Jie Fu

SourceMassive Editing for Large Language Models via Meta Learning

The paper measures editing success, generalization success, and locality success as the number of simultaneous edits scales from 2^5 to 2^14 on three released LLMs of increasing size. The authors explicitly note that 'as the model scale increases, the performance of all editing methods improve.' On GPT-J, MALMEN maintains ES of 99.7% and GS of 92.8% at 8192 edits, while on GPT-2 at 512 edits the same method achieves ES 99.7% and GS 96.1%. Locality also improves: LS rises from 16.5% on GPT-2 (m=512) to 25.1% on GPT-J (m=8192). The observation is made in the context of a specific editing method but is stated to hold across all methods tested.

Evidence
correlational
Key metric
GPT-2 (124M) m=512: ES 99.7%, GS 96.1%, LS 16.5%; GPT-J (6B) m=8192: ES 99.7%, GS 92.8%, LS 25.1%; 'As the model scale increases, the performance of all editing methods improve'; 'malmen can edit more than two orders of magnitude more facts than mend on gpt-j with similar performance'
Caveat
The three models have different architectures (decoder-only GPT-2, encoder-decoder T5-XL, decoder-only GPT-J), so the scale trend is confounded with architecture. The comparison is also at different numbers of edits (512 vs 8192), and the paper does not isolate model scale from architecture as a variable.
Model
GPT-2, T5 T5-XL, GPT-J
Concepts
Scale-dependent behaviour
Datasets
ZSRE [eval], Natural Questions / NaturalQA [source]
Methods
MEND [compared-to], MEMIT [compared-to]
Related work
MEND [compared-to], MEMIT [compared-to]
Extraction
automatic-extraction